Place: Large Lecture Room – CVC
Affiliation: CiSRA, Canon Information Systems Research Australia
Measuring image quality is a complex process that often requires elements of subjective analysis in order to be reliable when the final judge is a human observer. Preference studies show that slight variations in colour have drastically different outcomes in quality perception depending on where they occur. A few D E of difference in the colour of a wall may have no influence, while a shift of a single D E on a face will have a major impact.
The perceived quality of images does not in general depend on the image as a whole, but on a few salient regions within. Saliency in images is an essential to identify which parts of an image are of particular importance to human observers for subsequent processing. Consequently, a number of algorithms have been developed, that purport to automatically predict an images salient regions.
Most saliency algorithms are based on either the physiology of the human visual system or on general image statistics, and are designed with a broad scope in mind. However, studies of human attention and eye movement show that visual attention maps can vary significantly depending on the task, due to the influence of cognitive processes usually not taken into consideration by algorithms.
Identifying important regions for perceptual image quality measurement being a critical task, we devise an experimental framework to obtain visual attention maps and compare these to the saliency maps predicted by state-of-the-art algorithms. Measures of correlation and precision-recall curves indicate that automatic saliency measurement is not much better than random, and far from the performance of observers, perhaps suggesting that image quality assessment has more to do with high-level cognitive processes than with low-level vision.